Lomax, Anthony

In this paper we present and discuss the performance of the procedure for earthquake location and characterization implemented in the Italian Candidate Tsunami Service Provider at the Istituto Nazionale di Geofisica e Vulcanologia (INGV) in Rome. Following the ICG/NEAMTWS guidelines, the first tsunami warning messages are based only on seismic information, i.e., epicenter location, hypocenter depth, and magnitude, which are automatically computed by the software Early-est. Early-est is a package for rapid location and seismic/tsunamigenic characterization of earthquakes. The Early-est software package operates using offline-event or continuous-real-time seismic waveform data to perform trace processing and picking, and, at a regular report interval, phase association, event detection, hypocenter location, and event characterization. Early-est also provides mb, Mwp, and Mwpd magnitude estimations. mb magnitudes are preferred for events with Mwp ≲ 5.8, while Mwpd estimations are valid for events with Mwp ≳ 7.2. In this paper we present the earthquake parameters computed by Early-est between the beginning of March 2012 and the end of December 2014 on a global scale for events with magnitude M ≥ 5.5, and we also present the detection timeline. We compare the earthquake parameters automatically computed by Early-est with the same parameters listed in reference catalogs. Such reference catalogs are manually revised/verified by scientists. The goal of this work is to test the accuracy and reliability of the fully automatic locations provided by Early-est. In our analysis, the epicenter location, hypocenter depth and magnitude parameters do not differ significantly from the values in the reference catalogs. Both mb and Mwp magnitudes show differences to the reference catalogs. We thus derived correction functions in order to minimize the differences and correct biases between our values and the ones from the reference catalogs. Correction of the Mwp distance dependency is particularly relevant, since this magnitude refers to the larger and probably tsunamigenic earthquakes. Mwp values at stations with epicentral distance Δ ≲ 30° are significantly overestimated with respect to the CMT-global solutions, whereas Mwp values at stations with epicentral distance Δ ≳ 90° are slightly underestimated. After applying such distance correction the Mwp provided by Early-est differs from CMT-global catalog values of about δ Mwp ≈ 0.0 ∓ 0.2. Early-est continuously acquires time-series data and updates the earthquake source parameters. Our analysis shows that the epicenter coordinates and the magnitude values converge within less than 10 min (5 min in the Mediterranean region) toward the stable values. Our analysis shows that we can compute Mwp magnitudes that do not display short epicentral distance dependency overestimation, and we can provide robust and reliable earthquake source parameters to compile tsunami warning messages within less than 15 min after the event origin time.

After an earthquake, rapid, real-time assessment of hazards such as ground shaking and tsunami potential is important for early warning and emergency response. Tsunami potential depends on seafloor displacement, which is related to the length, L, width, W, mean slip, D, and depth, z, of earthquake rupture. Currently, the primary discriminant for tsunami potential is the centroid-moment tensor magnitude, MCMT, representing the seismic potency LWD, and w estimated through an indirect, inversion procedure. The obtained MCMT and the implied LWD w value vary with the depth of faulting, assumed earth model and other factors, and is only available 30 min or more after an earthquake. The use of more direct procedures for hazard assessment, when available, could avoid these problems and aid in effective early warning. Here we present a direct procedure for rapid assessment of earthquake tsunami potential using two, simple measures on P-wave seismograms—the dominant period on the velocity records, Td, and the likelihood that the high-frequency, apparent rupture-duration, T0, exceeds 50–55 s. T0 can be related to the critical parameters L and z, while Td may be related to W, D or z. For a set of recent, large earthquakes, we show that the period-duration product T T gives more information on tsunami impact and size than MCMT and other currently used d0w discriminants. All discriminants have difficulty in assessing the tsunami potential for oceanic strike-slip and backarc or upper plate, intraplate earthquake types. Our analysis and results suggest that tsunami potential is not directly related to the potency LWD from the ‘seismic’ faulting model, as is assumed with the use of the MCMT discriminant. Instead, knowledge of w rupture length, L, and depth, z, alone can constrain well the tsunami potential of an earthquake, with explicit determination of fault width, W, and slip, D, being of secondary importance. With available real-time seismogram data, rapid calculation of the direct, period-duration discriminant can be completed within 6–10 min after an earthquake occurs and thus can aid in effective and reliable tsunami early warning.

In their comment, Yamomoto and co-authors are primarily concerned with the existence and effect of large values of minimum and maximum phase residuals in our analysis and locations using the 2014 observations, as listed in Tables S7 and S8 in the supplementary material of our paper (Batsi et al, 2018). We retain these large residuals in the tables and analysis since they have vanishingly small effect on the NonLinLoc locations, since the used, equal differential time (EDT) location algorithm (Lomax, 2008; Lomax et al., 2009) is highly robust to outlier readings. In the case of our Marmara study, phases with residuals larger than 1-2sec have near zero weight in the locations and corrected phase data. However, we agree the larger residuals may have had adverse effect on the generation of station corrections, though this, in turn, would also be mitigated by the robust location procedure. As a result, we consider that the location discrepancies between Yamomoto et al (2017) and Batsi et al. (2018) are not due to effects of excessively large residuals on the station corrections or locations. Instead, we propose that, as in many seismicity studies, error and uncertainty in the absolute hypocenter locations is primarily related to error in the velocity model and insufficient geometrical coverage of the source zones by the available seismic stations. To support this proposition, and following the recommendation of Yamamoto et al., we recalculate station corrections for our 2014 data set and then relocate the 14 common events (Table A) that were located by both Yamamoto et al. (2017) and ourselves (see Table 9 in Batsi et al., 2018, with correct Yamomoto’s location for event 3: 40.8058N, 27.9504E, 13.411km). We first generate station corrections as described in Batsi et al. (2018) using all events from 2014 which comply with the Batsi et al. (2018) location criteria (number of stations ≥ 5; number of phases ≥ 6; (3) root mean square (rms) location error ≤ 0.5s; azimuthal gap ≤ 180°), except that we explicitly exclude from the analysis any P or S residuals > 3.0s when generating station corrections (Table B). We then relocate in the high‐resolution, 3D, P‐velocity model, as described in Batsi et al. (2018), the 14 common events using these station corrections. Figure 1 shows, for the 14 common events listed I Table A, the absolute NonLinLoc maximum likelihood and expectation hypocenters, and location probability density (pdf) clouds for our absolute relocations, along with the corresponding Yamamoto et al. (2017) double-difference relocations and Batsi et (2018) relative (NonDiffLoc) locations. For sake of clarity, calculation results are detailed in Figure 2 for each individual event (1 to 14). The full information on the earthquake location spatial uncertainty is shown by the pdf clouds, while the maximum-likelihood hypocenter is the best solution point and the expectation hypocenter shows a weighted mean or “center of mass” of the cloud. The pdf clouds show a large uncertainty in hypocenter depth, the formal standard error in depth ranges from 2-9km. There is also a large separation between the maximum likelihood and expectation hypocenters for some events. These results underline the large uncertainty in depth determination and corresponding instability in any one-point measure chosen as a hypocenter. However, despite these uncertainties and instabilities, the Yamamoto et al. (2017) hypocenters remain generally deeper than the maximum likelihood and expectation hypocenters for our relocations, positioned towards the deeper uncertainty limits of our locations (e.g. the lower portion of the pdf clouds), and the Yamamoto et al. (2017) epicenters fall near the Main Marmara fault (MMF) while our relocated epicenters define off axis seismicity, along secondary faults from the MMF system. Thus our relocated events, which explicitly exclude excessively large residuals, still show differences with the Yamamoto et al. (2017) events, but not as large as those we found in our original study. Based on our recalculated NonLinLoc absolute locations, we suspect that  Yamamoto et al (2017) results are systematically too deep and Batsi et al (2018) systematically too shallow, compared to what should be expected. These differences in epicenter and depth, along with the size and shape of the pdf clouds for our relocations, are most easily explained by differences in the 3D velocity models and by differences in available stations and the consequent network geometry . However, while the epicentral distances at most of the OBS stations are shorter than the focal depths, as noted by Yamomoto et al., the elongation of our pdf clouds in depth suggests that an increase in network aperture with more distant stations, along with an accurate 3D model, is required to better constrain depth. High-resolution earthquake epicenter and depth determinations below the Sea of Marmara is a difficult problem, yet of critical importance. To better understand why the two studies produce different results, and to obtain the best possible locations, the best action is to increase the number of constraints by merging the two OBS datasets, and examine, step by step, the effects of locations methods, network geometry and 3D velocity models from the two studies. Sharing the data (or phase picks and model) would provide an unique opportunity to give real, direct insight into these issues. We suspect that epicenters will shift as a function of used velocity model and station set, and that in all cases depth uncertainty is large, as is clearly represented in the NonLinLoc location, pdf clouds, while linearized location error estimates usually show lower uncertainty.

Effective early warning, emergency response, and information dissemination for earthquakes and tsunamis require rapid characterization of an earthquake’s location, size, and other parameters, usually provided by real‐time seismogram analysis using established, rule‐based, seismological procedures. Powerful, new machine learning (ML) tools analyze basic data using little or no rule‐based knowledge, and an ML deep convolutional neural network (CNN) can operate directly on seismogram waveforms with little preprocessing and without feature extraction. How a CNN will perform for rapid automated earthquake detection and characterization using short single‐station waveforms is an issue of fundamental importance for earthquake monitoring. For an initial investigation of this issue, we adapt an existing CNN for local earthquake detection and epicentral classification using single‐station waveforms (Perol et al., 2018), to form a new CNN, ConvNetQuake_INGV, to characterize earthquakes at any distance (local to far‐teleseismic). ConvNetQuake_INGV operates directly on 50‐s three‐component broadband single‐station waveforms to detect seismic events and obtain binned probabilistic estimates of the distance, azimuth, depth, and magnitude of the event. The best performance of ConvNetQuake_INGV is obtained using a last convolutional layer with fewer nodes than the number of output classifications, a form of information bottleneck. We show that ConvNetQuake_INGV detects very well (accuracy 87%) and characterizes moderately well earthquakes over a broad range of distances and magnitudes, and we analyze outlier results and indications of overfitting of the CNN training data. We find weak evidence that the CNN is performing more than high‐dimensional regression and pattern recognition, and is generalizing information or learning, to provide useful characterization of new events not represented in the training data. We expect that real‐time ML procedures such as ConvNetQuake_INGV, perhaps incorporating rule‐based knowledge, will ultimately prove valuable for rapid detection and characterization of earthquakes for earthquake response and tsunami early warning.

Effective tsunami early warning for coastlines near a tsunamigenic earthquake requires notification within 5-15 minutes. We have shown recently that tsunamigenic earthquakes have an apparent rupture duration, T0, greater than about 50 s. Here we show that T0 gives more information on tsunami importance than moment magnitude, Mw, and we introduce a procedure using seismograms recorded near an earthquake to rapidly determine if T0 is likely to exceed T=50 or 100 s. We show that this “duration-exceedance” procedure can be completed within 3-10 min after the earthquake occurs, depending on station density, and that it correctly identifies most recent earthquakes which produced large or devastating tsunamis. This identification forms a complement to initial estimates of the location, depth and magnitude of an earthquake to improve the reliability of tsunami early warning, and, in some cases, may make possible such warning.

The 26 December 2004 M9 Sumatra–Andaman Islands earthquake caused a tsunami that devastated Indian Ocean coasts within three hours after the earthquake. Improved tsunami warning and emergency response for future great earthquakes require knowing an earthquake’s size within minutes after the event. Although the hypocenter of a distant earthquake is routinely determined from the fi rst seismic P waves within about 15 min, several hours may pass before a reliable size determination for very large earthquakes is available (e.g., for the 2004 Sumatra–Andaman earthquake see Menke and Levin [2005]).

This study describes a deep convolutional neural network (CNN) based technique for the prediction of intensity measurements (IMs) of ground shaking. The input data to the CNN model consists of multistation 3C broadband and accelerometric waveforms recorded during the 2016 Central Italy earthquake sequence for M $\ge$ 3.0. We find that the CNN is capable of predicting accurately the IMs at stations far from the epicenter and that have not yet recorded the maximum ground shaking when using a 10 s window starting at the earthquake origin time. The CNN IM predictions do not require previous knowledge of the earthquake source (location and magnitude). Comparison between the CNN model predictions and the predictions obtained with Bindi et al. (2011) GMPE (which require location and magnitude) has shown that the CNN model features similar error variance but smaller bias. Although the technique is not strictly designed for earthquake early warning, we found that it can provide useful estimates of ground motions within 15-20 sec after earthquake origin time depending on various setup elements (e.g., times for data transmission, computation, latencies). The technique has been tested on raw data without any initial data pre-selection in order to closely replicate real-time data streaming. When noise examples were included with the earthquake data, the CNN was found to be stable predicting accurately the ground shaking intensity corresponding to the noise amplitude.

After an earthquake underwater, rapid real-time assessment of earthquake parameters is important for emergency response related to infrastructure damage and, perhaps more exigently, for issuing warnings of the possibility of an impending tsunami. Since 2005, the Istituto Nazionale di Geofisica e Vulcanologia (INGV) has worked on the rapid quantification of earthquake magnitude and tsunami potential, especially for the Mediterranean area. This work includes quantification of earthquake size from standard moment tensor inversion, quantification of earthquake size and tsunamigenic potential using P waveforms, and calculation of an archive of readily accessible tsunami scenarios.

We show that relative earthquake location using double-difference methods requires an accurate knowledge of the velocity structure throughout the study region to prevent artifacts in the relative position of hypocenters. The velocity structure determines the ray paths between hypocenters and receivers. These ray paths, and the corresponding ray take-off angles at the hypocenters, determine the partial derivatives of travel time with respect to the hypocentral coordinates which form the inversion kernel that maps double-differences into hypocentral perturbations. Thus the large-scale velocity structure enters into the core of the double-difference technique. By employing a 1D layered model with sharp interfaces to perform double-difference inversion of synthetic data generated using a simple, 1D gradient model; we show that inappropriate choice of the velocity model, combined with unbalanced source-receiver distributions, can lead to significant distortion and bias in the relative hypocenter positions of closely spaced events.

We present a duration-amplitude procedure for rapid determination of a moment magnitude, Mwpd, for large earthquakes using P-wave recordings at teleseismic distances. Mwpd can be obtained within 20 minutes or less after the event origin time as the required data is currently available in near-real time. The procedure determines apparent source durations, T0, from high-frequency, P-wave records, and estimates moments through integration of broadband displacement waveforms over the interval tP to tP+T0, where tP is the P arrival time. We apply the duration-amplitude methodology to 79 recent, large earthquakes (Global Centroid- Moment Tensor magnitude, MwCMT, 6.6 to 9.3) with diverse source types. The results show that a scaling of the moment estimates for interplate thrust and possibly tsunami earthquakes is necessary to best match MwCMT. With this scaling, Mwpd matches MwCMT typically within ±0.2 magnitude units, with a standard deviation of σ=0.11, equaling or outperforming other approaches to rapid magnitude determination. Furthermore, Mwpd does not exhibit saturation; that is, for the largest events, Mwpd does not systematically underestimate MwCMT. The obtained durations and duration-amplitude moments allow rapid estimation of an energy-to-moment parameter Θ* used for identification of tsunami earthquakes. Our results show that Θ* ≤ -5.7 is an appropriate cutoff for this identification, but also show that neither Θ* nor Mw is a good indicator for tsunamigenic events in general. For these events we find that a reliable indicator is simply that the duration T0 is greater than about 50 sec. The explicit use of the source duration for integration of displacement seismograms, the moment scaling, and other characteristics of the duration-amplitude methodology make it an extension of the widely used, Mwp, rapid-magnitude procedure. The need for a moment scaling for interplate thrust and possibly tsunami earthquakes may have important implications for the source physics of these events.